[Interim research report] Taking features out of superposition with sparse autoencoders — LessWrong

We're thankful for helpful comments from Trenton Bricken, Eric Winsor, Noa Nabeshima, and Sid Black. …

x [Interim research report] Taking features out of superposition with sparse autoencoders — LessWrong Sparse Autoencoders (SAEs) Interpretability (ML & AI) Superposition Conjecture (org) Bucket Errors AI Frontpage 156 [Interim research report] Taking features out of superposition with sparse autoencoders by Lee Sharkey , Dan Braun , beren 13th Dec 2022 AI Alignment Forum 27 min read 23 156 Ω 69 We're thankful for helpful comments from Trenton Bricken, Eric Winsor, Noa Nabeshima, and Sid Black. This post is part of the work done at Conjecture . TL;DR : Recent results from Anthropic suggest that

related reading

• Towards Monosemanticity: Decomposing Language Models With Dictionary Learningtransformer-circuits.pub
• Sparse Autoencoders Find Highly Interpretable Features in Language Modelsarxiv.org
• Toy Models of Superpositiontransformer-circuits.pub
• An Intuitive Explanation of Sparse Autoencoders for LLM Interpretability | Adam Karvonenadamkarvonen.github.io
• Open Source Replication & Commentary on Anthropic's Dictionary Learning Paper — LessWronglesswrong.com
• A gentle introduction to sparse autoencodersnickjiang.substack.com
• Do sparse autoencoders find "true features"? — LessWronglesswrong.com
• Circuits Updates - February 2024transformer-circuits.pub
• [2406.04093] Scaling and evaluating sparse autoencodersarxiv.org
• A gentle introduction to sparse autoencoders — LessWronglesswrong.com
• Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnettransformer-circuits.pub
• [2410.12101] The Persian Rug: solving toy models of superposition using large-scale symmetriesarxiv.org